Trust and Reputation for inferring quality of resources
Trust and Reputation for
inferring quality of resources
invited talk at First International Workshop on Quality Control in
Digital Libraries (QCDL'06)
2728 April 2006, Udine, Italy
ITC/iRST, Trento, Italy
Slides licenced under CreativeCommons AttributionShareAlike (see last slide for more info) 1
■ Recommender Systems
■ New trend: explicit trust and trust metrics
■ Local and Global Trust Metrics
■ Space for subjectivity? Experiments on real
community of Epinions.com
■ Risks: Tyranny of the majority / Daily Me
■ Suggestion in modelling online systems (for
example, digital libraries)
Recommender Systems in digital libraries
■ Recommender Systems suggest items the
user might like
0) Users express ratings (opinions)
1) RS find users similar to active user
2) RS recommends to active user items liked
Works for every domain: songs, movies, jokes, ..., digital 3
New trend: explicit trust
■ New trend: consider explicit trust between
Problems in computing user similarity > ask it
Users can express which other users they trust
And specifying the level of trust (i.e. In [0,1])
Concept used in Emarketplaces (Ebay.com,
Epinions.com, Amazon.com), News sites (Slashdot.org,
Kuro5hin.org), P2P networks (eDonkey, Gnutella, JXTA),
Jobs sites (LinkedIn, Ryze), Advogato.org, CouchSurfing, 4
■ Aggregate all the trust statements to produce a
trust network. A node is a user.
A direct edge is a trust statement
0.2 Properties of Trust:
0.9 weighted (0=distrust, 1=max trust)
ME Doc asymmetric contextdependent
Trust Metric (TM):
? ? Uses existing edges for predicting values
of trust for nonexisting edges,
Cory Mary thanks to trust propagation (if you trust
someone, then you have some degree of
trust in anyone that person trusts). 5
PageRank: a trust metric?
Imagine the web as a
■ Nodes are web pages,
Edges are links (not
Web Web weighted).
page page “importance” of every
single page based on
number and quality of
Web incoming edges...
■ So, YES: PageRank is
TM perspective: Local or Global
Mary Mena Bill
How much Bill can be trusted?
On average (by the community)?
ME 1 By Mary?
And by ME?
■ Global Trust Metrics:
“Reputation” of user is based on number and quality of incoming edges. Bill has just one
predicted trust value (0.5). pred_trust(Bill)=0.5
PageRank (eBay, Slashdot, ). Work bad for controversial people (bush)
■ Local Trust Metrics
Trust is subjective > consider personal views (trust “Bill”?)
AppleSeed, Golbeck TM, Advogato, ...
Local can be more effective if people are not standardized.
■ Most of the systems use global trust metrics (ebay,
google, slashdot, ...)
■ Most papers assume there are (globally agreed)
good peers (that gives correct ratings) and
malicious or wrong peers (that don't agree with
■ This assumption is not realistic (next slide)
■ It is dangerous: encourages herd behaviour and
penalizes creative thinkers, black sheeps,
Some evidence: Epinions.com
What is Epinions.com?
■ Community web site where users can
Write reviews about items and give them ratings
Express their Web of Trust (“Users whose reviews and
ratings you have consistently found to be valuable”)
Express their Block List (“Users whose reviews and ratings
... offensive, inaccurate, or in general not valuable”)
■ Reviews of TRUSTed users are more visible
■ Reviews of DISTRUSTed users are hidden
Evidence from real online community of 150.000 users).
Dr.P's Web of Trust
(Block List is hidden)
trust or distrust Dr.P?
Ratings given by Dr.P
Are there CORRECT ratings?
■ What is the correct rating of movie “Titanic”?
■ What is the correct rating of cd
■ ... what is the correct rating of “Divina
Commedia”? Of my paper?
■ IF 99% of people likes “Divina Commedia”,
I'm forced to like it? Otherwise I'm wrong or
■ No correct ratings. Just different subjective
Are there correct trust statements?
■ Intuitively: a Controversial User is
TRUSTED by some users
DISTRUSTED by some users
■ Do you want an example?
Controversial Users: an example
100M people 100M people
If you don't know Bush, should you trust Bush?
T(Bush)=0.5? Make sense? Here global metrics don't.
Some numbers about controversiality
■ Epinions.com dataset
Real Users: ~150K
Edges (Trust / Distrust): 841K (717K / 124K)
~85K received at least one judgement (trust or distrust)
17.090 (>20%) are at least 1controversial (at least 1
user disagrees with the majority) > Non negligible
1.247 are at least 10controversial
144 are at least 40controversial
1 user is 212controversial! (~400 trust her, 212 distrust
■ Comparing 2 metrics about accuracy in
Global: ebaylike. Trust(A)=#trust/(#trust+#distrust)
Local: MoleTrust, based on Trust Propagation from
current user (simple and fast)
Cycles are a problem > Order peers
based on distance from source user
Trust of users at level k is based only
on trust of users at level k1 (and k)
Trust propagation horizon & decay
■ How do we compare metrics?
■ Leaveoneout: Remove an edge in Trust
Network and try to predict it. Then compute
error as absolute difference between Real
and Predicted value.
Exp. on Controversiality Percentage
CP~0 = Controversial User
Error Ebay = 0.5 on
Ebay Controversiality percentage Controversial Users
Error MoleTrust2 smaller
but not as small as we
would like: can we reach 0?
MoleTrust2 Controversiality percentage 17
Controversial Users: an example
1 1 0
R 1 0
100M people 100M people
Local Metric makes more sense. Your trust in
Bush depends on your trusted users!
T(R,Bush)=1 T(D,Bush)=0 18
Controversiality Epinions: summary
■ Most papers assume a peer has a unique quality
value (there are good peers and bad peers, goal
is to spot bad)
■ IRREALISTIC assumption (Evidence from real
online community of 150,000 users).
■ Consequence: we need Local Trust Metrics
(personalized) [But most papers propose Global
■ Ref: [Controversial Users demand Local Trust Metrics: an
Experimental Study on Epinions.com Community, Massa,
Tyranny of the Majority
■ Resist the temptation to model your system
(i.e. digital library) with good and bad peers
because this is not realistic.
■ And it is dangerous
Tyranny of the majority [Democracy in America, de
Tocqueville, 1835] and [On Liberty, John Stuart Mill,
for one minority, which by definition has opinions that
are different from the ones of the majority, there is no
way to be protected “against the tyranny of the
prevailing opinion and feeling”. 20
Tyranny of the Majority: examples
Examples of minorities
■ ... basically evolution in society happens only if there is
someone who “thinks different”.
■ Suggestion: don't crush (or burn!) different thinkers!
Minority's opinions are opportunities.
■ No need to introduce it: online systems allow (local)
personalization ... however ...
Risk on the opposite side: DAILY ME
■ “Daily me” (aka “echo chambers”)
[Cass Sunstein. Republic.com. Princeton University
■ “technology has greatly increased people's ability to filter what
they want to read, see, and hear, to encounter only opinions of
like minded people and never again be confronted with people
with different ideas and opinions”
■ Risk: segmentation of society
The two extremes
Tyranny of the majority <> Daily
Is there a balance in the middle?
Could be no ratings at all? ... Wikipedia works
Why not stopping assigning “Nobel” prizes?
But what we teach to school pupils as “correct”?
Forced by law to discuss with people we don't
... this is more sociology than anything else ... 23
■ The more decentralized the enviroment (read
■ the more needed to rely on decentralized
assessment and control of quality (ratings,
trust > recommender systems, reputation
systems, trust metrics).
Community of peers selfmoderate itself
■ Resist temptation to design systems in which
there are concepts as “correct” and “wrong”
rating, “good” and “bad” user.
■ Offer both global and a local view.
(global is needed? Need for a nobel prize? What is
For example with a slidebar that lets you explore
the personalized view (daily me) and global view
(majority) and mixes of them. 25
Licence of these slides
These slides are released under
You are free:
* to copy, distribute, display, and perform the work
* to make derivative works
* to make commercial use of the work
Under the following conditions:
Attribution. You must attribute the work in the manner specified by the author or licensor.
Share Alike. If you alter, transform, or build upon this work, you may distribute the resulting work only under a license identical to
* For any reuse or distribution, you must make clear to others the license terms of this work.
* Any of these conditions can be waived if you get permission from the copyright holder.
Your fair use and other rights are in no way affected by the above.
More info at http://creativecommons.org/licenses/bysa/2.5/
Thanks for your attention